I implemented the core Transformer blocks from the ground up: encoder, decoder, multi-head attention, position-wise feed-forward network, and positional encodings. The goal was to understand the paper's mechanics by writing them in PyTorch rather than relying on a library.

The model is trained end-to-end for language translation, with attention weights visualized to interpret how the decoder focuses on source tokens. This project formed my deep-dive foundation before moving into modern LLMs and agentic systems.

Tech: PyTorch, NumPy, Python.
Repo: Language Translation using Transformers